Claude Fable 5: What Anthropic's Most Capable Model Means for Builders
Anthropic's Claude Fable 5 is the most capable model the company has released to the general public — and after an unusual 18-day interruption, it is available worldwide again as of July 1, 2026. If you build with large language models, here is what actually matters: what it does well, what it costs, the new behaviors you have to design around, and how to adopt it without surprises.
What Claude Fable 5 is
Released on June 9, 2026, Claude Fable 5 (claude-fable-5) is Anthropic's most capable widely released model, built for demanding reasoning and long-horizon, asynchronous agentic work — the kind of days-long, multi-step tasks earlier models could not sustain. It shares the exact same underlying weights as the restricted Claude Mythos 5; the only difference is that Fable 5 ships with active safety classifiers, while Mythos 5 (fewer guardrails) is limited to vetted partners under Project Glasswing.
The specs that matter
| Property | Claude Fable 5 |
|---|---|
| API model ID | claude-fable-5 |
| Context window | 1,000,000 tokens (default = max) |
| Max output | 128,000 tokens per request |
| Input price | $10 / million tokens |
| Output price | $50 / million tokens |
| Prompt-cache read | 90% off input |
| Thinking | Adaptive, always on |
| Data retention | 30 days; no Zero-Data-Retention option |
Two things stand out. First, the price is exactly double Claude Opus 4.8 ($5 / $25). Fable 5 is a premium tier — you reach for it when a task genuinely needs frontier capability, not for routine calls. Second, the 90% prompt-caching discount on input is the single biggest lever for keeping costs sane on long-context, agentic workloads.
The "effort" knob replaces thinking budgets
Fable 5 drops the old budget_tokens extended-thinking control. Instead you set an effort level: low, medium, high (default), xhigh, or max. This matters for your bill: cost scales steeply from high to xhigh to max, and many task classes plateau well before max. Anthropic reports that even low effort frequently beats the previous generation's xhigh. The practical takeaway: sweep effort per task class in your own evals instead of hardcoding max everywhere.
The behavior you must design around: refusals and fallback
The headline integration change is that Fable 5's safety classifiers — covering areas like cybersecurity, biology, and chemistry — can decline a request. When that happens, requests commonly fall back to Claude Opus 4.8, and blocked queries are billed at Opus rates. If you are shipping to real users, you need explicit refusal handling and a fallback model configured, or a subset of your traffic will fail in ways you did not plan for.
This is exactly where a unified, OpenAI-compatible gateway earns its keep: you can call claude-fable-5 and declare claude-opus-4-8 (or the cheaper claude-sonnet-5) as an automatic fallback behind a single API key, instead of hand-rolling retry logic against several provider SDKs. If that sounds useful, you can wire it up through one endpoint here.
Benchmarks, briefly
Anthropic positions Fable 5 as state of the art across coding, analytics, vision, and computer use: top of CursorBench and Cognition's FrontierBench for coding, the first model to break 90% on its core long-running analytics benchmark (a 10-point jump over Opus), the best result it has measured on its end-to-end "vibe-coding" ViBench, and the new leader on AutomationBench. Treat vendor benchmarks as directional — but the pattern is consistent: long-horizon, tool-heavy work is where Fable 5 separates from the pack.
Availability — and the fine print
After the US government applied export controls on June 12 that forced a full suspension, the controls were lifted on June 30 and Fable 5 returned globally on July 1 across the Claude platform, Claude.ai, Claude Code, and Cowork, with AWS, Google Cloud, and Microsoft Foundry being re-enabled. One caveat to plan around: on Pro, Max, Team, and select Enterprise plans, Fable 5 counts for up to 50% of your weekly usage limit through July 7, after which it moves to usage credits. Anthropic also shipped Claude Sonnet 5 ($3 / $15) — a cheaper model for running agents — alongside the restoration.
How to adopt it without surprises
- Reserve Fable 5 for the hard 10% — long-horizon agents, complex refactors, senior-level analysis — and route routine calls to Sonnet 5 or Opus 4.8.
- Turn on prompt caching; on repetitive long-context work the 90% input discount is decisive.
- Start at
lowormediumeffort and raise it only where your evals show a real gain. - Always configure a fallback for classifier refusals so a blocked request degrades gracefully instead of erroring.
- Watch the retention constraint: Fable 5 is not available under Zero-Data-Retention, which may matter for regulated workloads.
Bottom line
Claude Fable 5 is a genuine step up for ambitious, long-running work — at a premium price and with new refusal and fallback behavior you have to engineer for. The teams that win with it will be the ones that route deliberately: the right model for each task, caching on, effort tuned, fallback wired.
Want to try claude-fable-5 next to Opus 4.8 and Sonnet 5 behind one OpenAI-compatible key, with automatic fallback and unified billing? Create a key and make your first call.